Explore · 每周精选
发现最受关注的研究论文,追踪研究趋势,订阅感兴趣的期刊与关键词。
Hao Liang, Xiaochen Ma, Zhou Liu +32 authors
DataFlow is an LLM-driven data preparation framework that enhances data quality and reproducibility for various tasks, improving LLM performance with automatically generated pipelines.
50 篇论文 · 按点赞排序
Wanghan Xu, Yuhao Zhou, Yifan Zhou +104 authors
A framework for Scientific General Intelligence (SGI) is presented, evaluated using SGI-Bench, and improved with Test-Time Reinforcement Learning, highlighting gaps in existing models' scientific capabilities.
Jintao Zhang, Kaiwen Zheng, Kai Jiang +5 authors
TurboDiffusion accelerates video generation by 100-200x using attention acceleration, step distillation, and quantization, while maintaining video quality.
Jianhong Bai, Xiaoshi Wu, Xintao Wang +9 authors
SemanticGen addresses slow convergence and computational costs in video generation by using a two-stage diffusion model approach that first generates semantic features and then VAE latents, leading to faster convergence and high-quality results.
Chen Hu, Haikuo Du, Heng Wang +64 authors
Step-DeepResearch, an end-to-end agent enhanced with a data synthesis strategy and progressive training, achieves expert-level capabilities in deep research scenarios, outperforming established models.
Xiaopeng Lin, Shijie Lian, Bin Yu +8 authors
Proposed Egocentric2Embodiment pipeline translates human egocentric videos into structured training data for robots, enhancing their egocentric understanding and task performance.
Kelvin Li, Chuyi Shang, Leonid Karlinsky +3 authors
While Large Multimodal Models (LMMs) have made significant progress, they remain largely text-centric, relying on language as their core reasoning modality. As a result, they are limited in their ability to handle reasoning tasks that are predominantly visual. Recent approaches have sought to address this by supervising intermediate visual steps with helper images, depth maps, or image crops. However, these strategies impose restrictive priors on what "useful" visual abstractions look like, add heavy annotation costs, and struggle to generalize across tasks. To address this critical limitation, we propose a task-agnostic mechanism that trains LMMs to discover and use visual reasoning tokens without explicit supervision. These tokens attend globally and re-encode the image in a task-adaptive way, enabling the model to extract relevant visual information without hand-crafted supervision. Our approach outperforms direct fine-tuning and achieves state-of-the-art results on a diverse range of vision-centric tasks -- including those where intermediate abstractions are hard to specify -- while also generalizing to multi-task instruction tuning.
Jiaqi Tang, Jianmin Chen, Wei Wei +7 authors
A novel framework, Robust-R1, enhances multimodal large language models' robustness to visual degradations through explicit modeling, supervised fine-tuning, reward-driven alignment, and dynamic reasoning depth scaling, achieving state-of-the-art performance on real-world degradation benchmarks.
Weichen Fan, Haiwen Diao, Quan Wang +2 authors
Unified Autoencoding combines semantic and pixel-level information through a frequency-band modulator, resulting in a latent space with state-of-the-art performance on image benchmarks.
Yuqiao Tan, Minzheng Wang, Shizhu He +6 authors
The paper decomposes the policy of large language models into internal layer and modular policies, revealing distinct reasoning patterns across layers and proposing Bottom-up Policy Optimization to enhance performance on complex reasoning tasks.
Seijin Kobayashi, Yanick Schimpf, Maximilian Schlegel +12 authors
Large-scale autoregressive models pretrained on next-token prediction and finetuned with reinforcement learning (RL) have achieved unprecedented success on many problem domains. During RL, these models explore by generating new outputs, one token at a time. However, sampling actions token-by-token can result in highly inefficient learning, particularly when rewards are sparse. Here, we show that it is possible to overcome this problem by acting and exploring within the internal representations of an autoregressive model. Specifically, to discover temporally-abstract actions, we introduce a higher-order, non-causal sequence model whose outputs control the residual stream activations of a base autoregressive model. On grid world and MuJoCo-based tasks with hierarchical structure, we find that the higher-order model learns to compress long activation sequence chunks onto internal controllers. Critically, each controller executes a sequence of behaviorally meaningful actions that unfold over long timescales and are accompanied with a learned termination condition, such that composing multiple controllers over time leads to efficient exploration on novel tasks. We show that direct internal controller reinforcement, a process we term "internal RL", enables learning from sparse rewards in cases where standard RL finetuning fails. Our results demonstrate the benefits of latent action generation and reinforcement in autoregressive models, suggesting internal RL as a promising avenue for realizing hierarchical RL within foundation models.
Junyu Zhang, Yifan Sun, Tianang Leng +4 authors
A framework called Laws of Reasoning (LoRe) is introduced to theoretically define desired reasoning behaviors in Large Reasoning Models, with a focus on compute and accuracy laws, and a benchmark (LoRe-Bench) to measure these properties.
Runtao Liu, Ziyi Liu, Jiaqi Tang +4 authors
A multi-agent framework, involving a master LLM, grounding agent, and vision agent, enhances long-video QA by improving temporal grounding and leveraging visual and textual data.
Zhongwei Zhang, Fuchen Long, Wei Li +4 authors
ReCo is a novel instructional video editing paradigm that enhances accuracy and reduces token interference by incorporating constraint modeling and regularization techniques during in-context generation.
Jiangjie Chen, Wenxiang Chen, Jiacheng Du +19 authors
Seed-Prover 1.5, a formal theorem-proving model using large-scale agentic reinforcement learning and an efficient test-time scaling workflow, demonstrates superior performance in solving mathematical problems across various levels with reduced computational resources.
Shengchao Zhou, Yuxin Chen, Yuying Ge +4 authors
DSR Suite enhances vision-language models with dynamic spatial reasoning through automated data generation and a geometry selection module that integrates geometric priors.
Chiao-An Yang, Ryo Hachiuma, Sifei Liu +4 authors
4D-RGPT, a specialized multimodal LLM, enhances 4D perception in video inputs through Perceptual 4D Distillation and is evaluated on R4D-Bench, a new benchmark for depth-aware dynamic scenes.
NVIDIA, Aaron Blakeman, Aaron Grattafiori +355 authors
We introduce the Nemotron 3 family of models - Nano, Super, and Ultra. These models deliver strong agentic, reasoning, and conversational capabilities. The Nemotron 3 family uses a Mixture-of-Experts hybrid Mamba-Transformer architecture to provide best-in-class throughput and context lengths of up to 1M tokens. Super and Ultra models are trained with NVFP4 and incorporate LatentMoE, a novel approach that improves model quality. The two larger models also include MTP layers for faster text generation. All Nemotron 3 models are post-trained using multi-environment reinforcement learning enabling reasoning, multi-step tool use, and support granular reasoning budget control. Nano, the smallest model, outperforms comparable models in accuracy while remaining extremely cost-efficient for inference. Super is optimized for collaborative agents and high-volume workloads such as IT ticket automation. Ultra, the largest model, provides state-of-the-art accuracy and reasoning performance. Nano is released together with its technical report and this white paper, while Super and Ultra will follow in the coming months. We will openly release the model weights, pre- and post-training software, recipes, and all data for which we hold redistribution rights.
NVIDIA, Aaron Blakeman, Aaron Grattafiori +310 authors
We present Nemotron 3 Nano 30B-A3B, a Mixture-of-Experts hybrid Mamba-Transformer language model. Nemotron 3 Nano was pretrained on 25 trillion text tokens, including more than 3 trillion new unique tokens over Nemotron 2, followed by supervised fine tuning and large-scale RL on diverse environments. Nemotron 3 Nano achieves better accuracy than our previous generation Nemotron 2 Nano while activating less than half of the parameters per forward pass. It achieves up to 3.3x higher inference throughput than similarly-sized open models like GPT-OSS-20B and Qwen3-30B-A3B-Thinking-2507, while also being more accurate on popular benchmarks. Nemotron 3 Nano demonstrates enhanced agentic, reasoning, and chat abilities and supports context lengths up to 1M tokens. We release both our pretrained Nemotron 3 Nano 30B-A3B Base and post-trained Nemotron 3 Nano 30B-A3B checkpoints on Hugging Face.
Yuxi Xiao, Longfei Li, Shen Yan +5 authors
SpatialTree, a cognitive-science-inspired hierarchy, evaluates and improves spatial abilities in MLLMs across multiple levels, revealing transfer dynamics and proposing an auto-think strategy for consistent performance enhancement.
Jiongxiao Wang, Qiaojing Yan, Yawei Wang +6 authors
A novel RL framework, SAGE, enhances LLM-based agents' self-improvement capabilities by systematically incorporating skills from a skill library, leading to better performance and efficiency in new environments.
Shilong Zhang, He Zhang, Zhifei Zhang +11 authors
Latent diffusion models using representation encoder features face challenges in semantic compactness and pixel-level reconstruction, which are addressed through a semantic-pixel reconstruction objective that enables compact yet semantically rich representations for unified text-to-image and image editing tasks.
Jiawei Liu, Junqiao Li, Jiangfan Deng +11 authors
The "one-shot" technique represents a distinct and sophisticated aesthetic in filmmaking. However, its practical realization is often hindered by prohibitive costs and complex real-world constraints. Although emerging video generation models offer a virtual alternative, existing approaches typically rely on naive clip concatenation, which frequently fails to maintain visual smoothness and temporal coherence. In this paper, we introduce DreaMontage, a comprehensive framework designed for arbitrary frame-guided generation, capable of synthesizing seamless, expressive, and long-duration one-shot videos from diverse user-provided inputs. To achieve this, we address the challenge through three primary dimensions. (i) We integrate a lightweight intermediate-conditioning mechanism into the DiT architecture. By employing an Adaptive Tuning strategy that effectively leverages base training data, we unlock robust arbitrary-frame control capabilities. (ii) To enhance visual fidelity and cinematic expressiveness, we curate a high-quality dataset and implement a Visual Expression SFT stage. In addressing critical issues such as subject motion rationality and transition smoothness, we apply a Tailored DPO scheme, which significantly improves the success rate and usability of the generated content. (iii) To facilitate the production of extended sequences, we design a Segment-wise Auto-Regressive (SAR) inference strategy that operates in a memory-efficient manner. Extensive experiments demonstrate that our approach achieves visually striking and seamlessly coherent one-shot effects while maintaining computational efficiency, empowering users to transform fragmented visual materials into vivid, cohesive one-shot cinematic experiences.
Yuanning Feng, Sinan Wang, Zhengxiang Cheng +2 authors
Sage is a human-free evaluation suite for LLM-as-a-Judge, using rational choice theory to assess local and global consistency, revealing significant reliability issues with current LLM judges.
Jinjing Zhao, Fangyun Wei, Zhening Liu +3 authors
Spatia, a spatial memory-aware video generation framework, maintains long-term spatial and temporal consistency by preserving and updating a 3D scene point cloud, enabling realistic video generation and interactive editing.
Hanyang Kong, Xingyi Yang, Xiaoxu Zheng +1 authors
WorldWarp addresses the challenge of generating consistent long-range videos by integrating a 3D geometric cache with a spatio-temporal diffusion model, ensuring structural consistency and textural refinement.
Dehai Min, Kailin Zhang, Tongtong Wu +1 authors
QuCo-RAG uses objective corpus statistics to mitigate hallucinations in large language models during generation, improving accuracy across various benchmarks.
Li-Zhong Szu-Tu, Ting-Lin Wu, Chia-Jui Chang +2 authors
Vision-language models exhibit significant popularity bias in architectural recognition, performing poorly on less-known structures despite high accuracy on famous buildings.
Guibin Zhang, Haotian Ren, Chong Zhan +5 authors
MemEvolve, a meta-evolutionary framework, enhances self-evolving memory systems by jointly evolving agents' experiential knowledge and memory architecture, leading to improved performance and generalization.
Zeyuan Allen-Zhu
Controlled synthetic pretraining reveals CANON LAYERS, lightweight architectural components that enhance information flow and model capabilities across various sequence architectures.
北京市昌平区探索星信息技术及软件开发工作室
京ICP备2026059466号